In versions 3 and 4 I show the model I treated on Google COLAB for this notebook https://www.kaggle.com/aikhmelnytskyy/bagging-rainforest
IMPORTANTLY! This notebook didn't work after the changes to kaggle, but thanks to a discussion by Martin Görner and Allohvk (https://www.kaggle.com/c/rfcx-species-audio-detection/discussion/216408 ), I made the necessary changes in version 4 and now everything works. Here are the changes: from
@tf.function
def _preprocess_img(x, training=False, test=False):
to
@tf.function def _preprocess_img(x, training=False, test=False):
And from
def _specaugment(image):
image = tfa.image.cutout(image, [HEIGHT, xsize[0]], offset=[HEIGHT//2, xoff[0]])
image = tfa.image.cutout(image, [HEIGHT, xsize[1]], offset=[HEIGHT//2, xoff[1]])
image = tfa.image.cutout(image, [ysize[0], WIDTH], offset=[yoff[0], WIDTH//2])
image = tfa.image.cutout(image, [ysize[1], WIDTH], offset=[yoff[1], WIDTH//2])
image = tf.squeeze(image, axis=0)
return image
to
#image = tfa.image.cutout(image, [HEIGHT, xsize[0]], offset=[HEIGHT//2, xoff[0]])
#image = tfa.image.cutout(image, [HEIGHT, xsize[1]], offset=[HEIGHT//2, xoff[1]])
#image = tfa.image.cutout(image, [ysize[0], WIDTH], offset=[yoff[0], WIDTH//2])
#image = tfa.image.cutout(image, [ysize[1], WIDTH], offset=[yoff[1], WIDTH//2])
image = tf.squeeze(image, axis=0)
return imageVersion 5 changes as shown in this discussion https://www.kaggle.com/c/rfcx-species-audio-detection/discussion/218930 (special thanks to the author)
I used these notebooks as a basis: https://www.kaggle.com/mekhdigakhramanian/rfcx-resnet50-tpu https://www.kaggle.com/khoongweihao/resnet34-more-augmentations-mixup-tta-inference
It is important to work with colab you need kaggle.json (https://www.kaggle.com/docs/api)
I also created a folder called Models on my Google Drive and put the kaggle.json file in it.
import os
COLAB=False
models_path=''
if not os.path.exists('../input/rfcx-species-audio-detection'):# Let's check if we use kaggle notebook
COLAB=True
import gc
from google.colab import drive
drive.mount('/content/drive')# You must grant COLAB access to your Google Drive
#!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
GCS_DS_PATH = 'gs://kds-5c677f76ce55440722b2a474a5492faa70847c05a8f5d722f5a37feb'#
#This is a path to a dataset that changes over time, so you need to constantly update it. To update the path just run the code:
#GCS_DS_PATH = KaggleDatasets (). Get_gcs_path ()
#print (GCS_DS_PATH)
models_path='/content/drive/MyDrive/Models/'# I created a folder called Models on my Google Drive and put the kaggle.json file in it
else:
from kaggle_datasets import KaggleDatasets
GCS_DS_PATH = KaggleDatasets().get_gcs_path('rfcx-species-audio-detection')
print (GCS_DS_PATH)
gs://kds-a7dd8b09d52950714bee1818d843af117accc125d4d289d9afcebfab
if COLAB:# Prepare the kaggle.json file for use
from google.colab import files
if not os.path.exists('/.kaggle/kaggle.json'):
!mkdir ~/.kaggle
if not os.path.exists('/content/drive/My Drive/Models/kaggle.json'):
files.upload()
!cp kaggle.json ~/.kaggle/
else:
!cp '/content/drive/My Drive/Models/kaggle.json' ~/.kaggle/
!chmod 600 ~/.kaggle/kaggle.json
if COLAB:# force TF to 2.2
!pip install -q tensorflow~=2.2.0 tensorflow_gcs_config~=2.2.0
import tensorflow as tf
import requests
import os
resp = requests.post("http://{}:8475/requestversion/{}".format(os.environ["COLAB_TPU_ADDR"].split(":")[0], tf.__version__))
if resp.status_code != 200:
print("Failed to switch the TPU to TF {}".format(version))
if COLAB: #%tensorflow_version 2.x
import tensorflow as tf
print("Tensorflow version " + tf.__version__)
try:
tpu = tf.distribute.cluster_resolver.TPUClusterResolver() # TPU detection
print('Running on TPU ', tpu.cluster_spec().as_dict()['worker'])
except ValueError:
raise BaseException('ERROR: Not connected to a TPU runtime; please see the previous cell in this notebook for instructions!')
tf.config.experimental_connect_to_cluster(tpu)
tf.tpu.experimental.initialize_tpu_system(tpu)
tpu_strategy = tf.distribute.experimental.TPUStrategy(tpu)
#tpu_strategy = tf.distribute.TPUStrategy(tpu)
!pip install -q tensorflow_io
import tensorflow_io as tfio
import tensorflow as tf
import gc
!pip install image-classifiers
!pip install tensorflow_addons==0.10.0
#0.11.2
import tensorflow_addons as tfa
#import tfa as tfa
import numpy as np
from pathlib import Path
import io
import matplotlib.pyplot as plt
!pip install soundfile
import soundfile as sf
import librosa
#!pip install kaggle_datasets
#from kaggle_datasets import KaggleDatasets
from tqdm import tqdm
import pandas as pd
from sklearn.model_selection import StratifiedKFold
import seaborn as sns
from IPython.display import Audio
from classification_models.keras import Classifiers
tf.__version__
WARNING: You are using pip version 21.0; however, version 21.0.1 is available. You should consider upgrading via the '/opt/conda/bin/python3.7 -m pip install --upgrade pip' command. Collecting image-classifiers Downloading image_classifiers-1.0.0-py3-none-any.whl (19 kB) Collecting keras-applications<=1.0.8,>=1.0.7 Downloading Keras_Applications-1.0.8-py3-none-any.whl (50 kB) |████████████████████████████████| 50 kB 1.5 MB/s Requirement already satisfied: numpy>=1.9.1 in /opt/conda/lib/python3.7/site-packages (from keras-applications<=1.0.8,>=1.0.7->image-classifiers) (1.19.5) Requirement already satisfied: h5py in /opt/conda/lib/python3.7/site-packages (from keras-applications<=1.0.8,>=1.0.7->image-classifiers) (2.10.0) Requirement already satisfied: six in /opt/conda/lib/python3.7/site-packages (from h5py->keras-applications<=1.0.8,>=1.0.7->image-classifiers) (1.15.0) Installing collected packages: keras-applications, image-classifiers Successfully installed image-classifiers-1.0.0 keras-applications-1.0.8 WARNING: You are using pip version 21.0; however, version 21.0.1 is available. You should consider upgrading via the '/opt/conda/bin/python3.7 -m pip install --upgrade pip' command. Collecting tensorflow_addons==0.10.0 Downloading tensorflow_addons-0.10.0-cp37-cp37m-manylinux2010_x86_64.whl (1.0 MB) |████████████████████████████████| 1.0 MB 2.9 MB/s Requirement already satisfied: typeguard>=2.7 in /opt/conda/lib/python3.7/site-packages (from tensorflow_addons==0.10.0) (2.10.0) Installing collected packages: tensorflow-addons Attempting uninstall: tensorflow-addons Found existing installation: tensorflow-addons 0.12.0 Uninstalling tensorflow-addons-0.12.0: Successfully uninstalled tensorflow-addons-0.12.0 Successfully installed tensorflow-addons-0.10.0 WARNING: You are using pip version 21.0; however, version 21.0.1 is available. You should consider upgrading via the '/opt/conda/bin/python3.7 -m pip install --upgrade pip' command.
/opt/conda/lib/python3.7/site-packages/tensorflow_addons/utils/ensure_tf_install.py:68: UserWarning: Tensorflow Addons supports using Python ops for all Tensorflow versions above or equal to 2.2.0 and strictly below 2.3.0 (nightly versions are not supported). The versions of TensorFlow you are currently using is 2.4.0 and is not supported. Some things might work, some things might not. If you were to encounter a bug, do not file an issue. If you want to make sure you're using a tested and supported configuration, either change the TensorFlow version or the TensorFlow Addons's version. You can find the compatibility matrix in TensorFlow Addon's readme: https://github.com/tensorflow/addons UserWarning,
Requirement already satisfied: soundfile in /opt/conda/lib/python3.7/site-packages (0.10.3.post1)
Requirement already satisfied: cffi>=1.0 in /opt/conda/lib/python3.7/site-packages (from soundfile) (1.14.4)
Requirement already satisfied: pycparser in /opt/conda/lib/python3.7/site-packages (from cffi>=1.0->soundfile) (2.20)
WARNING: You are using pip version 21.0; however, version 21.0.1 is available.
You should consider upgrading via the '/opt/conda/bin/python3.7 -m pip install --upgrade pip' command.
'2.4.0'
from classification_models.keras import Classifiers
#SEED = 42
import random
SEED = random.randint(0, 10000)# !!!!
def seed_everything(seed):
random.seed(seed)
os.environ['PYTHONHASHSEED'] = str(seed)
np.random.seed(seed)
tf.random.set_seed(seed)
seed_everything(SEED)
# from https://github.com/qubvel/classification_models
ResNet34, preprocess_input = Classifiers.get('resnet34')
cfg = {
'parse_params': {
'cut_time': 10,
},
'data_params': {
'sample_time': 6, # assert 60 % sample_time == 0
'spec_fmax': 24000.0,
'spec_fmin': 40.0,
'spec_mel': 300,
'mel_power': 2,
'img_shape': (300, 670)
},
'model_params': {
'batchsize_per_tpu': 8,
'iteration_per_epoch': 128,
'epoch': 25,# 1 epoch just for example
'arch': ResNet34,
'arch_preprocess': preprocess_input,
'freeze_to': 0, # Freeze to backbone.layers[:freeze_to]. If None, all layers in the backbone will be freezed.
'loss': {
'fn': tfa.losses.SigmoidFocalCrossEntropy,
'params': {},
},
'optim': {
'fn': tfa.optimizers.RectifiedAdam,
'params': {'lr': 2e-3, 'total_steps': 18*64, 'warmup_proportion': 0.3, 'min_lr': 1e-6},
},
'mixup': True # False
}
}
# detect and init the TPU
if not COLAB:
tpu = tf.distribute.cluster_resolver.TPUClusterResolver()
tf.config.experimental_connect_to_cluster(tpu)
tf.tpu.experimental.initialize_tpu_system(tpu)
print("All devices: ", tf.config.list_logical_devices('TPU'))
tpu_strategy = tf.distribute.experimental.TPUStrategy(tpu)
All devices: [LogicalDevice(name='/job:worker/replica:0/task:0/device:TPU:5', device_type='TPU'), LogicalDevice(name='/job:worker/replica:0/task:0/device:TPU:4', device_type='TPU'), LogicalDevice(name='/job:worker/replica:0/task:0/device:TPU:3', device_type='TPU'), LogicalDevice(name='/job:worker/replica:0/task:0/device:TPU:2', device_type='TPU'), LogicalDevice(name='/job:worker/replica:0/task:0/device:TPU:6', device_type='TPU'), LogicalDevice(name='/job:worker/replica:0/task:0/device:TPU:7', device_type='TPU'), LogicalDevice(name='/job:worker/replica:0/task:0/device:TPU:1', device_type='TPU'), LogicalDevice(name='/job:worker/replica:0/task:0/device:TPU:0', device_type='TPU')]
AUTOTUNE = tf.data.experimental.AUTOTUNE
TRAIN_TFREC = GCS_DS_PATH + "/tfrecords/train"
TEST_TFREC = GCS_DS_PATH + "/tfrecords/test"
CUT = cfg['parse_params']['cut_time']
SR = 48000 # all wave's sample rate may be 48k
TIME = cfg['data_params']['sample_time']
FMAX = cfg['data_params']['spec_fmax']
FMIN = cfg['data_params']['spec_fmin']
N_MEL = cfg['data_params']['spec_mel']
HEIGHT, WIDTH = cfg['data_params']['img_shape']
CLASS_N = 24
raw_dataset = tf.data.TFRecordDataset([TRAIN_TFREC + '/00-148.tfrec'])
raw_dataset
<TFRecordDatasetV2 shapes: (), types: tf.string>
feature_description = {
'recording_id': tf.io.FixedLenFeature([], tf.string, default_value=''),
'audio_wav': tf.io.FixedLenFeature([], tf.string, default_value=''),
'label_info': tf.io.FixedLenFeature([], tf.string, default_value=''),
}
parse_dtype = {
'audio_wav': tf.float32,
'recording_id': tf.string,
'species_id': tf.int32,
'songtype_id': tf.int32,
't_min': tf.float32,
'f_min': tf.float32,
't_max': tf.float32,
'f_max':tf.float32,
'is_tp': tf.int32
}
@tf.function
def _parse_function(example_proto):
sample = tf.io.parse_single_example(example_proto, feature_description)
wav, _ = tf.audio.decode_wav(sample['audio_wav'], desired_channels=1) # mono
label_info = tf.strings.split(sample['label_info'], sep='"')[1]
labels = tf.strings.split(label_info, sep=';')
@tf.function
def _cut_audio(label):
items = tf.strings.split(label, sep=',')
spid = tf.squeeze(tf.strings.to_number(items[0], tf.int32))
soid = tf.squeeze(tf.strings.to_number(items[1], tf.int32))
tmin = tf.squeeze(tf.strings.to_number(items[2]))
fmin = tf.squeeze(tf.strings.to_number(items[3]))
tmax = tf.squeeze(tf.strings.to_number(items[4]))
fmax = tf.squeeze(tf.strings.to_number(items[5]))
tp = tf.squeeze(tf.strings.to_number(items[6], tf.int32))
tmax_s = tmax * tf.cast(SR, tf.float32)
tmin_s = tmin * tf.cast(SR, tf.float32)
cut_s = tf.cast(CUT * SR, tf.float32)
all_s = tf.cast(60 * SR, tf.float32)
tsize_s = tmax_s - tmin_s
cut_min = tf.cast(
tf.maximum(0.0,
tf.minimum(tmin_s - (cut_s - tsize_s) / 2,
tf.minimum(tmax_s + (cut_s - tsize_s) / 2, all_s) - cut_s)
), tf.int32
)
cut_max = cut_min + CUT * SR
_sample = {
'audio_wav': tf.reshape(wav[cut_min:cut_max], [CUT*SR]),
'recording_id': sample['recording_id'],
'species_id': spid,
'songtype_id': soid,
't_min': tmin - tf.cast(cut_min, tf.float32)/tf.cast(SR, tf.float32),
'f_min': fmin,
't_max': tmax - tf.cast(cut_min, tf.float32)/tf.cast(SR, tf.float32),
'f_max': fmax,
'is_tp': tp
}
return _sample
samples = tf.map_fn(_cut_audio, labels, dtype=parse_dtype)
return samples
parsed_dataset = raw_dataset.map(_parse_function).unbatch()
@tf.function
def _cut_wav(x):
# random cut in training
cut_min = tf.random.uniform([], maxval=(CUT-TIME)*SR, dtype=tf.int32)
cut_max = cut_min + TIME * SR
cutwave = tf.reshape(x['audio_wav'][cut_min:cut_max], [TIME*SR])
y = {}
y.update(x)
y['audio_wav'] = cutwave
y['t_min'] = tf.maximum(0.0, x['t_min'] - tf.cast(cut_min, tf.float32) / SR)
y['t_max'] = tf.maximum(0.0, x['t_max'] - tf.cast(cut_min, tf.float32) / SR)
return y
@tf.function
def _cut_wav_val(x):
# center crop in validation
cut_min = (CUT-TIME)*SR // 2
cut_max = cut_min + TIME * SR
cutwave = tf.reshape(x['audio_wav'][cut_min:cut_max], [TIME*SR])
y = {}
y.update(x)
y['audio_wav'] = cutwave
y['t_min'] = tf.maximum(0.0, x['t_min'] - tf.cast(cut_min, tf.float32) / SR)
y['t_max'] = tf.maximum(0.0, x['t_max'] - tf.cast(cut_min, tf.float32) / SR)
return y
@tf.function
def _filtTP(x):
return x['is_tp'] == 1
def show_wav(sample, ax):
wav = sample["audio_wav"].numpy()
rate = SR
ax.plot(np.arange(len(wav)) / rate, wav)
ax.set_title(
sample["recording_id"].numpy().decode()
+ ("/%d" % sample["species_id"])
+ ("TP" if sample["is_tp"] else "FP"))
return Audio((wav * 2**15).astype(np.int16), rate=rate)
fig, ax = plt.subplots(figsize=(15, 3))
show_wav(next(iter(parsed_dataset)), ax)
@tf.function
def _wav_to_spec(x):
mel_power = cfg['data_params']['mel_power']
stfts = tf.signal.stft(x["audio_wav"], frame_length=2048, frame_step=512, fft_length=2048)
spectrograms = tf.abs(stfts) ** mel_power
# Warp the linear scale spectrograms into the mel-scale.
num_spectrogram_bins = stfts.shape[-1]
lower_edge_hertz, upper_edge_hertz, num_mel_bins = FMIN, FMAX, N_MEL
linear_to_mel_weight_matrix = tf.signal.linear_to_mel_weight_matrix(
num_mel_bins, num_spectrogram_bins, SR, lower_edge_hertz,
upper_edge_hertz)
mel_spectrograms = tf.tensordot(
spectrograms, linear_to_mel_weight_matrix, 1)
mel_spectrograms.set_shape(spectrograms.shape[:-1].concatenate(
linear_to_mel_weight_matrix.shape[-1:]))
# Compute a stabilized log to get log-magnitude mel-scale spectrograms.
log_mel_spectrograms = tf.math.log(mel_spectrograms + 1e-6)
y = {
'audio_spec': tf.transpose(log_mel_spectrograms), # (num_mel_bins, frames)
}
y.update(x)
return y
spec_dataset = parsed_dataset.filter(_filtTP).map(_cut_wav).map(_wav_to_spec)
plt.figure(figsize=(12,5))
for i, s in enumerate(spec_dataset.take(3)):
plt.subplot(1,3,i+1)
plt.imshow(s['audio_spec'])
plt.show()
import librosa.display
import matplotlib.patches as patches
def show_spectrogram(sample, ax, showlabel=False):
S_dB = sample["audio_spec"].numpy()
img = librosa.display.specshow(S_dB, x_axis='time',
y_axis='mel', sr=SR,
fmax=FMAX, fmin=FMIN, ax=ax, cmap='magma')
ax.set(title=f'Mel-frequency spectrogram of {sample["recording_id"].numpy().decode()}')
sid, fmin, fmax, tmin, tmax, istp = (
sample["species_id"], sample["f_min"], sample["f_max"], sample["t_min"], sample["t_max"], sample["is_tp"])
ec = '#00ff00' if istp == 1 else '#0000ff'
ax.add_patch(
patches.Rectangle(xy=(tmin, fmin), width=tmax-tmin, height=fmax-fmin, ec=ec, fill=False)
)
if showlabel:
ax.text(tmin, fmax,
f"{sid.numpy().item()} {'tp' if istp == 1 else 'fp'}",
horizontalalignment='left', verticalalignment='bottom', color=ec, fontsize=16)
fig, ax = plt.subplots(figsize=(15,3))
show_spectrogram(next(iter(spec_dataset)), ax, showlabel=True)
/opt/conda/lib/python3.7/site-packages/librosa/display.py:974: MatplotlibDeprecationWarning: The 'basey' parameter of __init__() has been renamed 'base' since Matplotlib 3.3; support for the old name will be dropped two minor releases later. scaler(mode, **kwargs) /opt/conda/lib/python3.7/site-packages/librosa/display.py:974: MatplotlibDeprecationWarning: The 'linthreshy' parameter of __init__() has been renamed 'linthresh' since Matplotlib 3.3; support for the old name will be dropped two minor releases later. scaler(mode, **kwargs)
# in validation, annotations will come to the center
fig, ax = plt.subplots(figsize=(15,3))
show_spectrogram(next(iter(parsed_dataset.filter(_filtTP).map(_cut_wav_val).map(_wav_to_spec))), ax, showlabel=True)
for sample in spec_dataset.take(5):
fig, ax = plt.subplots(figsize=(15,3))
show_spectrogram(sample, ax, showlabel=True)
@tf.function
def _create_annot(x):
targ = tf.one_hot(x["species_id"], CLASS_N, on_value=x["is_tp"], off_value=0)
return {
'input': x["audio_spec"],
'target': tf.cast(targ, tf.float32)
}
annot_dataset = spec_dataset.map(_create_annot)
#@tf.function
def _preprocess_img(x, training=False, test=False):
image = tf.expand_dims(x, axis=-1)
image = tf.image.resize(image, [HEIGHT, WIDTH])
image = tf.image.per_image_standardization(image)
@tf.function
def _specaugment(image):
ERASE_TIME = 50
ERASE_MEL = 16
image = tf.squeeze(image, axis=2)
image = tfio.experimental.audio.time_mask(image, param=ERASE_TIME)
image = tfio.experimental.audio.freq_mask(image, param=ERASE_MEL)
image = tf.expand_dims(image, axis=2)
return image
if training:
# gaussian
gau = tf.keras.layers.GaussianNoise(0.3)
image = tf.cond(tf.random.uniform([]) < 0.5, lambda: gau(image, training=True), lambda: image)
# brightness
image = tf.image.random_brightness(image, 0.2)
# random left right flip (NEW)
image = tf.image.random_flip_left_right(image)
# specaugment
image = tf.cond(tf.random.uniform([]) < 0.5, lambda: _specaugment(image), lambda: image)
if test:
# specaugment
image = tf.cond(tf.random.uniform([]) < 0.5, lambda: _specaugment(image), lambda: image)
image = (image - tf.reduce_min(image)) / (tf.reduce_max(image) - tf.reduce_min(image)) * 255.0 # rescale to [0, 255]
image = tf.image.grayscale_to_rgb(image)
image = cfg['model_params']['arch_preprocess'](image)
return image
@tf.function
def _preprocess(x):
image = _preprocess_img(x['input'], training=True, test=False)
return (image, x["target"])
@tf.function
def _preprocess_val(x):
image = _preprocess_img(x['input'], training=False, test=False)
return (image, x["target"])
@tf.function
def _preprocess_test(x):
image = _preprocess_img(x['audio_spec'], training=False, test=True)
return (image, x["recording_id"])
for inp, targ in annot_dataset.map(_preprocess).take(2):
plt.imshow(inp.numpy()[:,:,0])
t = targ.numpy()
if t.sum() == 0:
plt.title(f'FP')
else:
plt.title(f'{t.nonzero()[0]}')
plt.colorbar()
plt.show()
from tensorflow.keras.layers import *
from tensorflow.keras import losses, models, optimizers
from tensorflow.keras.optimizers import Adam
def create_model():
#with strategy.scope():
#backbone = cfg['model_params']['arch'](include_top=False, weights='imagenet')
def Classifier(shape_):
backbone = cfg['model_params']['arch']((shape_), include_top=False, weights='imagenet')
def cbr(x, out_layer, kernel, stride, dilation):
x = Conv2D(out_layer, kernel_size=kernel, dilation_rate=dilation, strides=stride, padding="same")(x)
x = BatchNormalization()(x)
x = Activation("relu")(x)
return x
def wave_block(x, filters, kernel_size, n):
dilation_rates = [2**i for i in range(n)]
x = Conv2D(filters = filters,
kernel_size = 1,
padding = 'same')(x)
res_x = x
for dilation_rate in dilation_rates:
tanh_out = Conv2D(filters = filters,
kernel_size = kernel_size,
padding = 'same',
activation = 'tanh',
dilation_rate = dilation_rate)(x)
sigm_out = Conv2D(filters = filters,
kernel_size = kernel_size,
padding = 'same',
activation = 'sigmoid',
dilation_rate = dilation_rate)(x)
x = Multiply()([tanh_out, sigm_out])
x = Conv2D(filters = filters,
kernel_size = 1,
padding = 'same')(x)
res_x = Add()([res_x, x])
return res_x
#out1
def wavenet(layer):
x = cbr(layer, 192, 7, 1, 1)
x = BatchNormalization()(x)
x = wave_block(x, 192, 3, 1)
x = cbr(x, 96, 7, 1, 1)
x = BatchNormalization()(x)
x = wave_block(x, 96, 3, 1)
x = cbr(x, 48, 5, 1, 1)
x = BatchNormalization()(x)
x = wave_block(x, 48, 3, 1)
return x
def wavenet1(layer):
x = cbr(layer, 4, 7, 1, 1)
x = BatchNormalization()(x)
x = wave_block(x, 3, 3, 1)
x = cbr(x, 3, 7, 1, 1)
x = BatchNormalization()(x)
x = wave_block(x, 16, 3, 1)
x = cbr(x, 3, 5, 1, 1)
return x
#x = BatchNormalization()(x)
x0 = backbone#model
print('1')
#backbone.summary()
x1 = tf.keras.layers.GlobalAveragePooling2D()(x0.layers[-1].output) #-3,-7,-9,-15 for EF5
#x2 = tf.keras.layers.GlobalAveragePooling2D()(x0.layers[-3].output) # 803,799,797,791 for EF7
x3 = tf.keras.layers.GlobalAveragePooling2D()(x0.layers[-7].output)
#x4 = tf.keras.layers.GlobalAveragePooling2D()(x0.layers[-12].output)
x5 = tf.keras.layers.GlobalAveragePooling2D()(x0.layers[-18].output)
print('2')
x1=wavenet(x0.layers[-1].output)
x3=wavenet(x0.layers[-7].output)
x5=wavenet(x0.layers[-18].output)
x1 = tf.keras.layers.GlobalAveragePooling2D()(x1)
x3 = tf.keras.layers.GlobalAveragePooling2D()(x3)
x5 = tf.keras.layers.GlobalAveragePooling2D()(x5)
print('4')
#x = tf.concat([x1,x2,x3,x4,x5],axis = 1)
x = tf.concat([x1,x3,x5],axis = 1)
x = tf.keras.layers.Dropout(0.7)(x)
x = tf.keras.layers.Dense(192)(x)
#x = tf.keras.layers.BatchNormalization()(x)
x = tf.keras.layers.Dropout(0.4)(x)
#x = margin([x , label])
output = tf.keras.layers.Softmax(dtype='float32')(x)
output =tf.keras.layers.Dense(CLASS_N)(x)
print('5')
model = tf.keras.models.Model(inputs = x0.input, outputs = output)
#model.compile(optimizer=optimizer, loss=loss_fn, metrics=[LWLRAP(CLASS_N)])
return model
return Classifier([HEIGHT,WIDTH,3])
model = create_model()
model.summary()
Downloading data from https://github.com/qubvel/classification_models/releases/download/0.0.1/resnet34_imagenet_1000_no_top.h5
85524480/85521592 [==============================] - 1s 0us/step
1
2
4
5
Model: "model_1"
__________________________________________________________________________________________________
Layer (type) Output Shape Param # Connected to
==================================================================================================
data (InputLayer) [(None, 300, 670, 3) 0
__________________________________________________________________________________________________
bn_data (BatchNormalization) (None, 300, 670, 3) 9 data[0][0]
__________________________________________________________________________________________________
zero_padding2d (ZeroPadding2D) (None, 306, 676, 3) 0 bn_data[0][0]
__________________________________________________________________________________________________
conv0 (Conv2D) (None, 150, 335, 64) 9408 zero_padding2d[0][0]
__________________________________________________________________________________________________
bn0 (BatchNormalization) (None, 150, 335, 64) 256 conv0[0][0]
__________________________________________________________________________________________________
relu0 (Activation) (None, 150, 335, 64) 0 bn0[0][0]
__________________________________________________________________________________________________
zero_padding2d_1 (ZeroPadding2D (None, 152, 337, 64) 0 relu0[0][0]
__________________________________________________________________________________________________
pooling0 (MaxPooling2D) (None, 75, 168, 64) 0 zero_padding2d_1[0][0]
__________________________________________________________________________________________________
stage1_unit1_bn1 (BatchNormaliz (None, 75, 168, 64) 256 pooling0[0][0]
__________________________________________________________________________________________________
stage1_unit1_relu1 (Activation) (None, 75, 168, 64) 0 stage1_unit1_bn1[0][0]
__________________________________________________________________________________________________
zero_padding2d_2 (ZeroPadding2D (None, 77, 170, 64) 0 stage1_unit1_relu1[0][0]
__________________________________________________________________________________________________
stage1_unit1_conv1 (Conv2D) (None, 75, 168, 64) 36864 zero_padding2d_2[0][0]
__________________________________________________________________________________________________
stage1_unit1_bn2 (BatchNormaliz (None, 75, 168, 64) 256 stage1_unit1_conv1[0][0]
__________________________________________________________________________________________________
stage1_unit1_relu2 (Activation) (None, 75, 168, 64) 0 stage1_unit1_bn2[0][0]
__________________________________________________________________________________________________
zero_padding2d_3 (ZeroPadding2D (None, 77, 170, 64) 0 stage1_unit1_relu2[0][0]
__________________________________________________________________________________________________
stage1_unit1_conv2 (Conv2D) (None, 75, 168, 64) 36864 zero_padding2d_3[0][0]
__________________________________________________________________________________________________
stage1_unit1_sc (Conv2D) (None, 75, 168, 64) 4096 stage1_unit1_relu1[0][0]
__________________________________________________________________________________________________
add (Add) (None, 75, 168, 64) 0 stage1_unit1_conv2[0][0]
stage1_unit1_sc[0][0]
__________________________________________________________________________________________________
stage1_unit2_bn1 (BatchNormaliz (None, 75, 168, 64) 256 add[0][0]
__________________________________________________________________________________________________
stage1_unit2_relu1 (Activation) (None, 75, 168, 64) 0 stage1_unit2_bn1[0][0]
__________________________________________________________________________________________________
zero_padding2d_4 (ZeroPadding2D (None, 77, 170, 64) 0 stage1_unit2_relu1[0][0]
__________________________________________________________________________________________________
stage1_unit2_conv1 (Conv2D) (None, 75, 168, 64) 36864 zero_padding2d_4[0][0]
__________________________________________________________________________________________________
stage1_unit2_bn2 (BatchNormaliz (None, 75, 168, 64) 256 stage1_unit2_conv1[0][0]
__________________________________________________________________________________________________
stage1_unit2_relu2 (Activation) (None, 75, 168, 64) 0 stage1_unit2_bn2[0][0]
__________________________________________________________________________________________________
zero_padding2d_5 (ZeroPadding2D (None, 77, 170, 64) 0 stage1_unit2_relu2[0][0]
__________________________________________________________________________________________________
stage1_unit2_conv2 (Conv2D) (None, 75, 168, 64) 36864 zero_padding2d_5[0][0]
__________________________________________________________________________________________________
add_1 (Add) (None, 75, 168, 64) 0 stage1_unit2_conv2[0][0]
add[0][0]
__________________________________________________________________________________________________
stage1_unit3_bn1 (BatchNormaliz (None, 75, 168, 64) 256 add_1[0][0]
__________________________________________________________________________________________________
stage1_unit3_relu1 (Activation) (None, 75, 168, 64) 0 stage1_unit3_bn1[0][0]
__________________________________________________________________________________________________
zero_padding2d_6 (ZeroPadding2D (None, 77, 170, 64) 0 stage1_unit3_relu1[0][0]
__________________________________________________________________________________________________
stage1_unit3_conv1 (Conv2D) (None, 75, 168, 64) 36864 zero_padding2d_6[0][0]
__________________________________________________________________________________________________
stage1_unit3_bn2 (BatchNormaliz (None, 75, 168, 64) 256 stage1_unit3_conv1[0][0]
__________________________________________________________________________________________________
stage1_unit3_relu2 (Activation) (None, 75, 168, 64) 0 stage1_unit3_bn2[0][0]
__________________________________________________________________________________________________
zero_padding2d_7 (ZeroPadding2D (None, 77, 170, 64) 0 stage1_unit3_relu2[0][0]
__________________________________________________________________________________________________
stage1_unit3_conv2 (Conv2D) (None, 75, 168, 64) 36864 zero_padding2d_7[0][0]
__________________________________________________________________________________________________
add_2 (Add) (None, 75, 168, 64) 0 stage1_unit3_conv2[0][0]
add_1[0][0]
__________________________________________________________________________________________________
stage2_unit1_bn1 (BatchNormaliz (None, 75, 168, 64) 256 add_2[0][0]
__________________________________________________________________________________________________
stage2_unit1_relu1 (Activation) (None, 75, 168, 64) 0 stage2_unit1_bn1[0][0]
__________________________________________________________________________________________________
zero_padding2d_8 (ZeroPadding2D (None, 77, 170, 64) 0 stage2_unit1_relu1[0][0]
__________________________________________________________________________________________________
stage2_unit1_conv1 (Conv2D) (None, 38, 84, 128) 73728 zero_padding2d_8[0][0]
__________________________________________________________________________________________________
stage2_unit1_bn2 (BatchNormaliz (None, 38, 84, 128) 512 stage2_unit1_conv1[0][0]
__________________________________________________________________________________________________
stage2_unit1_relu2 (Activation) (None, 38, 84, 128) 0 stage2_unit1_bn2[0][0]
__________________________________________________________________________________________________
zero_padding2d_9 (ZeroPadding2D (None, 40, 86, 128) 0 stage2_unit1_relu2[0][0]
__________________________________________________________________________________________________
stage2_unit1_conv2 (Conv2D) (None, 38, 84, 128) 147456 zero_padding2d_9[0][0]
__________________________________________________________________________________________________
stage2_unit1_sc (Conv2D) (None, 38, 84, 128) 8192 stage2_unit1_relu1[0][0]
__________________________________________________________________________________________________
add_3 (Add) (None, 38, 84, 128) 0 stage2_unit1_conv2[0][0]
stage2_unit1_sc[0][0]
__________________________________________________________________________________________________
stage2_unit2_bn1 (BatchNormaliz (None, 38, 84, 128) 512 add_3[0][0]
__________________________________________________________________________________________________
stage2_unit2_relu1 (Activation) (None, 38, 84, 128) 0 stage2_unit2_bn1[0][0]
__________________________________________________________________________________________________
zero_padding2d_10 (ZeroPadding2 (None, 40, 86, 128) 0 stage2_unit2_relu1[0][0]
__________________________________________________________________________________________________
stage2_unit2_conv1 (Conv2D) (None, 38, 84, 128) 147456 zero_padding2d_10[0][0]
__________________________________________________________________________________________________
stage2_unit2_bn2 (BatchNormaliz (None, 38, 84, 128) 512 stage2_unit2_conv1[0][0]
__________________________________________________________________________________________________
stage2_unit2_relu2 (Activation) (None, 38, 84, 128) 0 stage2_unit2_bn2[0][0]
__________________________________________________________________________________________________
zero_padding2d_11 (ZeroPadding2 (None, 40, 86, 128) 0 stage2_unit2_relu2[0][0]
__________________________________________________________________________________________________
stage2_unit2_conv2 (Conv2D) (None, 38, 84, 128) 147456 zero_padding2d_11[0][0]
__________________________________________________________________________________________________
add_4 (Add) (None, 38, 84, 128) 0 stage2_unit2_conv2[0][0]
add_3[0][0]
__________________________________________________________________________________________________
stage2_unit3_bn1 (BatchNormaliz (None, 38, 84, 128) 512 add_4[0][0]
__________________________________________________________________________________________________
stage2_unit3_relu1 (Activation) (None, 38, 84, 128) 0 stage2_unit3_bn1[0][0]
__________________________________________________________________________________________________
zero_padding2d_12 (ZeroPadding2 (None, 40, 86, 128) 0 stage2_unit3_relu1[0][0]
__________________________________________________________________________________________________
stage2_unit3_conv1 (Conv2D) (None, 38, 84, 128) 147456 zero_padding2d_12[0][0]
__________________________________________________________________________________________________
stage2_unit3_bn2 (BatchNormaliz (None, 38, 84, 128) 512 stage2_unit3_conv1[0][0]
__________________________________________________________________________________________________
stage2_unit3_relu2 (Activation) (None, 38, 84, 128) 0 stage2_unit3_bn2[0][0]
__________________________________________________________________________________________________
zero_padding2d_13 (ZeroPadding2 (None, 40, 86, 128) 0 stage2_unit3_relu2[0][0]
__________________________________________________________________________________________________
stage2_unit3_conv2 (Conv2D) (None, 38, 84, 128) 147456 zero_padding2d_13[0][0]
__________________________________________________________________________________________________
add_5 (Add) (None, 38, 84, 128) 0 stage2_unit3_conv2[0][0]
add_4[0][0]
__________________________________________________________________________________________________
stage2_unit4_bn1 (BatchNormaliz (None, 38, 84, 128) 512 add_5[0][0]
__________________________________________________________________________________________________
stage2_unit4_relu1 (Activation) (None, 38, 84, 128) 0 stage2_unit4_bn1[0][0]
__________________________________________________________________________________________________
zero_padding2d_14 (ZeroPadding2 (None, 40, 86, 128) 0 stage2_unit4_relu1[0][0]
__________________________________________________________________________________________________
stage2_unit4_conv1 (Conv2D) (None, 38, 84, 128) 147456 zero_padding2d_14[0][0]
__________________________________________________________________________________________________
stage2_unit4_bn2 (BatchNormaliz (None, 38, 84, 128) 512 stage2_unit4_conv1[0][0]
__________________________________________________________________________________________________
stage2_unit4_relu2 (Activation) (None, 38, 84, 128) 0 stage2_unit4_bn2[0][0]
__________________________________________________________________________________________________
zero_padding2d_15 (ZeroPadding2 (None, 40, 86, 128) 0 stage2_unit4_relu2[0][0]
__________________________________________________________________________________________________
stage2_unit4_conv2 (Conv2D) (None, 38, 84, 128) 147456 zero_padding2d_15[0][0]
__________________________________________________________________________________________________
add_6 (Add) (None, 38, 84, 128) 0 stage2_unit4_conv2[0][0]
add_5[0][0]
__________________________________________________________________________________________________
stage3_unit1_bn1 (BatchNormaliz (None, 38, 84, 128) 512 add_6[0][0]
__________________________________________________________________________________________________
stage3_unit1_relu1 (Activation) (None, 38, 84, 128) 0 stage3_unit1_bn1[0][0]
__________________________________________________________________________________________________
zero_padding2d_16 (ZeroPadding2 (None, 40, 86, 128) 0 stage3_unit1_relu1[0][0]
__________________________________________________________________________________________________
stage3_unit1_conv1 (Conv2D) (None, 19, 42, 256) 294912 zero_padding2d_16[0][0]
__________________________________________________________________________________________________
stage3_unit1_bn2 (BatchNormaliz (None, 19, 42, 256) 1024 stage3_unit1_conv1[0][0]
__________________________________________________________________________________________________
stage3_unit1_relu2 (Activation) (None, 19, 42, 256) 0 stage3_unit1_bn2[0][0]
__________________________________________________________________________________________________
zero_padding2d_17 (ZeroPadding2 (None, 21, 44, 256) 0 stage3_unit1_relu2[0][0]
__________________________________________________________________________________________________
stage3_unit1_conv2 (Conv2D) (None, 19, 42, 256) 589824 zero_padding2d_17[0][0]
__________________________________________________________________________________________________
stage3_unit1_sc (Conv2D) (None, 19, 42, 256) 32768 stage3_unit1_relu1[0][0]
__________________________________________________________________________________________________
add_7 (Add) (None, 19, 42, 256) 0 stage3_unit1_conv2[0][0]
stage3_unit1_sc[0][0]
__________________________________________________________________________________________________
stage3_unit2_bn1 (BatchNormaliz (None, 19, 42, 256) 1024 add_7[0][0]
__________________________________________________________________________________________________
stage3_unit2_relu1 (Activation) (None, 19, 42, 256) 0 stage3_unit2_bn1[0][0]
__________________________________________________________________________________________________
zero_padding2d_18 (ZeroPadding2 (None, 21, 44, 256) 0 stage3_unit2_relu1[0][0]
__________________________________________________________________________________________________
stage3_unit2_conv1 (Conv2D) (None, 19, 42, 256) 589824 zero_padding2d_18[0][0]
__________________________________________________________________________________________________
stage3_unit2_bn2 (BatchNormaliz (None, 19, 42, 256) 1024 stage3_unit2_conv1[0][0]
__________________________________________________________________________________________________
stage3_unit2_relu2 (Activation) (None, 19, 42, 256) 0 stage3_unit2_bn2[0][0]
__________________________________________________________________________________________________
zero_padding2d_19 (ZeroPadding2 (None, 21, 44, 256) 0 stage3_unit2_relu2[0][0]
__________________________________________________________________________________________________
stage3_unit2_conv2 (Conv2D) (None, 19, 42, 256) 589824 zero_padding2d_19[0][0]
__________________________________________________________________________________________________
add_8 (Add) (None, 19, 42, 256) 0 stage3_unit2_conv2[0][0]
add_7[0][0]
__________________________________________________________________________________________________
stage3_unit3_bn1 (BatchNormaliz (None, 19, 42, 256) 1024 add_8[0][0]
__________________________________________________________________________________________________
stage3_unit3_relu1 (Activation) (None, 19, 42, 256) 0 stage3_unit3_bn1[0][0]
__________________________________________________________________________________________________
zero_padding2d_20 (ZeroPadding2 (None, 21, 44, 256) 0 stage3_unit3_relu1[0][0]
__________________________________________________________________________________________________
stage3_unit3_conv1 (Conv2D) (None, 19, 42, 256) 589824 zero_padding2d_20[0][0]
__________________________________________________________________________________________________
stage3_unit3_bn2 (BatchNormaliz (None, 19, 42, 256) 1024 stage3_unit3_conv1[0][0]
__________________________________________________________________________________________________
stage3_unit3_relu2 (Activation) (None, 19, 42, 256) 0 stage3_unit3_bn2[0][0]
__________________________________________________________________________________________________
zero_padding2d_21 (ZeroPadding2 (None, 21, 44, 256) 0 stage3_unit3_relu2[0][0]
__________________________________________________________________________________________________
stage3_unit3_conv2 (Conv2D) (None, 19, 42, 256) 589824 zero_padding2d_21[0][0]
__________________________________________________________________________________________________
add_9 (Add) (None, 19, 42, 256) 0 stage3_unit3_conv2[0][0]
add_8[0][0]
__________________________________________________________________________________________________
stage3_unit4_bn1 (BatchNormaliz (None, 19, 42, 256) 1024 add_9[0][0]
__________________________________________________________________________________________________
stage3_unit4_relu1 (Activation) (None, 19, 42, 256) 0 stage3_unit4_bn1[0][0]
__________________________________________________________________________________________________
zero_padding2d_22 (ZeroPadding2 (None, 21, 44, 256) 0 stage3_unit4_relu1[0][0]
__________________________________________________________________________________________________
stage3_unit4_conv1 (Conv2D) (None, 19, 42, 256) 589824 zero_padding2d_22[0][0]
__________________________________________________________________________________________________
stage3_unit4_bn2 (BatchNormaliz (None, 19, 42, 256) 1024 stage3_unit4_conv1[0][0]
__________________________________________________________________________________________________
stage3_unit4_relu2 (Activation) (None, 19, 42, 256) 0 stage3_unit4_bn2[0][0]
__________________________________________________________________________________________________
zero_padding2d_23 (ZeroPadding2 (None, 21, 44, 256) 0 stage3_unit4_relu2[0][0]
__________________________________________________________________________________________________
stage3_unit4_conv2 (Conv2D) (None, 19, 42, 256) 589824 zero_padding2d_23[0][0]
__________________________________________________________________________________________________
add_10 (Add) (None, 19, 42, 256) 0 stage3_unit4_conv2[0][0]
add_9[0][0]
__________________________________________________________________________________________________
stage3_unit5_bn1 (BatchNormaliz (None, 19, 42, 256) 1024 add_10[0][0]
__________________________________________________________________________________________________
stage3_unit5_relu1 (Activation) (None, 19, 42, 256) 0 stage3_unit5_bn1[0][0]
__________________________________________________________________________________________________
zero_padding2d_24 (ZeroPadding2 (None, 21, 44, 256) 0 stage3_unit5_relu1[0][0]
__________________________________________________________________________________________________
stage3_unit5_conv1 (Conv2D) (None, 19, 42, 256) 589824 zero_padding2d_24[0][0]
__________________________________________________________________________________________________
stage3_unit5_bn2 (BatchNormaliz (None, 19, 42, 256) 1024 stage3_unit5_conv1[0][0]
__________________________________________________________________________________________________
stage3_unit5_relu2 (Activation) (None, 19, 42, 256) 0 stage3_unit5_bn2[0][0]
__________________________________________________________________________________________________
zero_padding2d_25 (ZeroPadding2 (None, 21, 44, 256) 0 stage3_unit5_relu2[0][0]
__________________________________________________________________________________________________
stage3_unit5_conv2 (Conv2D) (None, 19, 42, 256) 589824 zero_padding2d_25[0][0]
__________________________________________________________________________________________________
add_11 (Add) (None, 19, 42, 256) 0 stage3_unit5_conv2[0][0]
add_10[0][0]
__________________________________________________________________________________________________
stage3_unit6_bn1 (BatchNormaliz (None, 19, 42, 256) 1024 add_11[0][0]
__________________________________________________________________________________________________
stage3_unit6_relu1 (Activation) (None, 19, 42, 256) 0 stage3_unit6_bn1[0][0]
__________________________________________________________________________________________________
zero_padding2d_26 (ZeroPadding2 (None, 21, 44, 256) 0 stage3_unit6_relu1[0][0]
__________________________________________________________________________________________________
stage3_unit6_conv1 (Conv2D) (None, 19, 42, 256) 589824 zero_padding2d_26[0][0]
__________________________________________________________________________________________________
stage3_unit6_bn2 (BatchNormaliz (None, 19, 42, 256) 1024 stage3_unit6_conv1[0][0]
__________________________________________________________________________________________________
stage3_unit6_relu2 (Activation) (None, 19, 42, 256) 0 stage3_unit6_bn2[0][0]
__________________________________________________________________________________________________
zero_padding2d_27 (ZeroPadding2 (None, 21, 44, 256) 0 stage3_unit6_relu2[0][0]
__________________________________________________________________________________________________
stage3_unit6_conv2 (Conv2D) (None, 19, 42, 256) 589824 zero_padding2d_27[0][0]
__________________________________________________________________________________________________
add_12 (Add) (None, 19, 42, 256) 0 stage3_unit6_conv2[0][0]
add_11[0][0]
__________________________________________________________________________________________________
stage4_unit1_bn1 (BatchNormaliz (None, 19, 42, 256) 1024 add_12[0][0]
__________________________________________________________________________________________________
stage4_unit1_relu1 (Activation) (None, 19, 42, 256) 0 stage4_unit1_bn1[0][0]
__________________________________________________________________________________________________
zero_padding2d_28 (ZeroPadding2 (None, 21, 44, 256) 0 stage4_unit1_relu1[0][0]
__________________________________________________________________________________________________
stage4_unit1_conv1 (Conv2D) (None, 10, 21, 512) 1179648 zero_padding2d_28[0][0]
__________________________________________________________________________________________________
stage4_unit1_bn2 (BatchNormaliz (None, 10, 21, 512) 2048 stage4_unit1_conv1[0][0]
__________________________________________________________________________________________________
stage4_unit1_relu2 (Activation) (None, 10, 21, 512) 0 stage4_unit1_bn2[0][0]
__________________________________________________________________________________________________
zero_padding2d_29 (ZeroPadding2 (None, 12, 23, 512) 0 stage4_unit1_relu2[0][0]
__________________________________________________________________________________________________
stage4_unit1_conv2 (Conv2D) (None, 10, 21, 512) 2359296 zero_padding2d_29[0][0]
__________________________________________________________________________________________________
stage4_unit1_sc (Conv2D) (None, 10, 21, 512) 131072 stage4_unit1_relu1[0][0]
__________________________________________________________________________________________________
add_13 (Add) (None, 10, 21, 512) 0 stage4_unit1_conv2[0][0]
stage4_unit1_sc[0][0]
__________________________________________________________________________________________________
stage4_unit2_bn1 (BatchNormaliz (None, 10, 21, 512) 2048 add_13[0][0]
__________________________________________________________________________________________________
stage4_unit2_relu1 (Activation) (None, 10, 21, 512) 0 stage4_unit2_bn1[0][0]
__________________________________________________________________________________________________
zero_padding2d_30 (ZeroPadding2 (None, 12, 23, 512) 0 stage4_unit2_relu1[0][0]
__________________________________________________________________________________________________
stage4_unit2_conv1 (Conv2D) (None, 10, 21, 512) 2359296 zero_padding2d_30[0][0]
__________________________________________________________________________________________________
stage4_unit2_bn2 (BatchNormaliz (None, 10, 21, 512) 2048 stage4_unit2_conv1[0][0]
__________________________________________________________________________________________________
stage4_unit2_relu2 (Activation) (None, 10, 21, 512) 0 stage4_unit2_bn2[0][0]
__________________________________________________________________________________________________
zero_padding2d_31 (ZeroPadding2 (None, 12, 23, 512) 0 stage4_unit2_relu2[0][0]
__________________________________________________________________________________________________
stage4_unit2_conv2 (Conv2D) (None, 10, 21, 512) 2359296 zero_padding2d_31[0][0]
__________________________________________________________________________________________________
add_14 (Add) (None, 10, 21, 512) 0 stage4_unit2_conv2[0][0]
add_13[0][0]
__________________________________________________________________________________________________
stage4_unit3_bn1 (BatchNormaliz (None, 10, 21, 512) 2048 add_14[0][0]
__________________________________________________________________________________________________
stage4_unit3_relu1 (Activation) (None, 10, 21, 512) 0 stage4_unit3_bn1[0][0]
__________________________________________________________________________________________________
zero_padding2d_32 (ZeroPadding2 (None, 12, 23, 512) 0 stage4_unit3_relu1[0][0]
__________________________________________________________________________________________________
stage4_unit3_conv1 (Conv2D) (None, 10, 21, 512) 2359296 zero_padding2d_32[0][0]
__________________________________________________________________________________________________
stage4_unit3_bn2 (BatchNormaliz (None, 10, 21, 512) 2048 stage4_unit3_conv1[0][0]
__________________________________________________________________________________________________
stage4_unit3_relu2 (Activation) (None, 10, 21, 512) 0 stage4_unit3_bn2[0][0]
__________________________________________________________________________________________________
zero_padding2d_33 (ZeroPadding2 (None, 12, 23, 512) 0 stage4_unit3_relu2[0][0]
__________________________________________________________________________________________________
stage4_unit3_conv2 (Conv2D) (None, 10, 21, 512) 2359296 zero_padding2d_33[0][0]
__________________________________________________________________________________________________
add_15 (Add) (None, 10, 21, 512) 0 stage4_unit3_conv2[0][0]
add_14[0][0]
__________________________________________________________________________________________________
bn1 (BatchNormalization) (None, 10, 21, 512) 2048 add_15[0][0]
__________________________________________________________________________________________________
relu1 (Activation) (None, 10, 21, 512) 0 bn1[0][0]
__________________________________________________________________________________________________
conv2d (Conv2D) (None, 10, 21, 192) 4817088 relu1[0][0]
__________________________________________________________________________________________________
conv2d_15 (Conv2D) (None, 10, 21, 192) 4817088 stage4_unit3_bn2[0][0]
__________________________________________________________________________________________________
conv2d_30 (Conv2D) (None, 12, 23, 192) 4817088 zero_padding2d_30[0][0]
__________________________________________________________________________________________________
batch_normalization (BatchNorma (None, 10, 21, 192) 768 conv2d[0][0]
__________________________________________________________________________________________________
batch_normalization_6 (BatchNor (None, 10, 21, 192) 768 conv2d_15[0][0]
__________________________________________________________________________________________________
batch_normalization_12 (BatchNo (None, 12, 23, 192) 768 conv2d_30[0][0]
__________________________________________________________________________________________________
activation (Activation) (None, 10, 21, 192) 0 batch_normalization[0][0]
__________________________________________________________________________________________________
activation_3 (Activation) (None, 10, 21, 192) 0 batch_normalization_6[0][0]
__________________________________________________________________________________________________
activation_6 (Activation) (None, 12, 23, 192) 0 batch_normalization_12[0][0]
__________________________________________________________________________________________________
batch_normalization_1 (BatchNor (None, 10, 21, 192) 768 activation[0][0]
__________________________________________________________________________________________________
batch_normalization_7 (BatchNor (None, 10, 21, 192) 768 activation_3[0][0]
__________________________________________________________________________________________________
batch_normalization_13 (BatchNo (None, 12, 23, 192) 768 activation_6[0][0]
__________________________________________________________________________________________________
conv2d_1 (Conv2D) (None, 10, 21, 192) 37056 batch_normalization_1[0][0]
__________________________________________________________________________________________________
conv2d_16 (Conv2D) (None, 10, 21, 192) 37056 batch_normalization_7[0][0]
__________________________________________________________________________________________________
conv2d_31 (Conv2D) (None, 12, 23, 192) 37056 batch_normalization_13[0][0]
__________________________________________________________________________________________________
conv2d_2 (Conv2D) (None, 10, 21, 192) 331968 conv2d_1[0][0]
__________________________________________________________________________________________________
conv2d_3 (Conv2D) (None, 10, 21, 192) 331968 conv2d_1[0][0]
__________________________________________________________________________________________________
conv2d_17 (Conv2D) (None, 10, 21, 192) 331968 conv2d_16[0][0]
__________________________________________________________________________________________________
conv2d_18 (Conv2D) (None, 10, 21, 192) 331968 conv2d_16[0][0]
__________________________________________________________________________________________________
conv2d_32 (Conv2D) (None, 12, 23, 192) 331968 conv2d_31[0][0]
__________________________________________________________________________________________________
conv2d_33 (Conv2D) (None, 12, 23, 192) 331968 conv2d_31[0][0]
__________________________________________________________________________________________________
multiply (Multiply) (None, 10, 21, 192) 0 conv2d_2[0][0]
conv2d_3[0][0]
__________________________________________________________________________________________________
multiply_3 (Multiply) (None, 10, 21, 192) 0 conv2d_17[0][0]
conv2d_18[0][0]
__________________________________________________________________________________________________
multiply_6 (Multiply) (None, 12, 23, 192) 0 conv2d_32[0][0]
conv2d_33[0][0]
__________________________________________________________________________________________________
conv2d_4 (Conv2D) (None, 10, 21, 192) 37056 multiply[0][0]
__________________________________________________________________________________________________
conv2d_19 (Conv2D) (None, 10, 21, 192) 37056 multiply_3[0][0]
__________________________________________________________________________________________________
conv2d_34 (Conv2D) (None, 12, 23, 192) 37056 multiply_6[0][0]
__________________________________________________________________________________________________
add_16 (Add) (None, 10, 21, 192) 0 conv2d_1[0][0]
conv2d_4[0][0]
__________________________________________________________________________________________________
add_19 (Add) (None, 10, 21, 192) 0 conv2d_16[0][0]
conv2d_19[0][0]
__________________________________________________________________________________________________
add_22 (Add) (None, 12, 23, 192) 0 conv2d_31[0][0]
conv2d_34[0][0]
__________________________________________________________________________________________________
conv2d_5 (Conv2D) (None, 10, 21, 96) 903264 add_16[0][0]
__________________________________________________________________________________________________
conv2d_20 (Conv2D) (None, 10, 21, 96) 903264 add_19[0][0]
__________________________________________________________________________________________________
conv2d_35 (Conv2D) (None, 12, 23, 96) 903264 add_22[0][0]
__________________________________________________________________________________________________
batch_normalization_2 (BatchNor (None, 10, 21, 96) 384 conv2d_5[0][0]
__________________________________________________________________________________________________
batch_normalization_8 (BatchNor (None, 10, 21, 96) 384 conv2d_20[0][0]
__________________________________________________________________________________________________
batch_normalization_14 (BatchNo (None, 12, 23, 96) 384 conv2d_35[0][0]
__________________________________________________________________________________________________
activation_1 (Activation) (None, 10, 21, 96) 0 batch_normalization_2[0][0]
__________________________________________________________________________________________________
activation_4 (Activation) (None, 10, 21, 96) 0 batch_normalization_8[0][0]
__________________________________________________________________________________________________
activation_7 (Activation) (None, 12, 23, 96) 0 batch_normalization_14[0][0]
__________________________________________________________________________________________________
batch_normalization_3 (BatchNor (None, 10, 21, 96) 384 activation_1[0][0]
__________________________________________________________________________________________________
batch_normalization_9 (BatchNor (None, 10, 21, 96) 384 activation_4[0][0]
__________________________________________________________________________________________________
batch_normalization_15 (BatchNo (None, 12, 23, 96) 384 activation_7[0][0]
__________________________________________________________________________________________________
conv2d_6 (Conv2D) (None, 10, 21, 96) 9312 batch_normalization_3[0][0]
__________________________________________________________________________________________________
conv2d_21 (Conv2D) (None, 10, 21, 96) 9312 batch_normalization_9[0][0]
__________________________________________________________________________________________________
conv2d_36 (Conv2D) (None, 12, 23, 96) 9312 batch_normalization_15[0][0]
__________________________________________________________________________________________________
conv2d_7 (Conv2D) (None, 10, 21, 96) 83040 conv2d_6[0][0]
__________________________________________________________________________________________________
conv2d_8 (Conv2D) (None, 10, 21, 96) 83040 conv2d_6[0][0]
__________________________________________________________________________________________________
conv2d_22 (Conv2D) (None, 10, 21, 96) 83040 conv2d_21[0][0]
__________________________________________________________________________________________________
conv2d_23 (Conv2D) (None, 10, 21, 96) 83040 conv2d_21[0][0]
__________________________________________________________________________________________________
conv2d_37 (Conv2D) (None, 12, 23, 96) 83040 conv2d_36[0][0]
__________________________________________________________________________________________________
conv2d_38 (Conv2D) (None, 12, 23, 96) 83040 conv2d_36[0][0]
__________________________________________________________________________________________________
multiply_1 (Multiply) (None, 10, 21, 96) 0 conv2d_7[0][0]
conv2d_8[0][0]
__________________________________________________________________________________________________
multiply_4 (Multiply) (None, 10, 21, 96) 0 conv2d_22[0][0]
conv2d_23[0][0]
__________________________________________________________________________________________________
multiply_7 (Multiply) (None, 12, 23, 96) 0 conv2d_37[0][0]
conv2d_38[0][0]
__________________________________________________________________________________________________
conv2d_9 (Conv2D) (None, 10, 21, 96) 9312 multiply_1[0][0]
__________________________________________________________________________________________________
conv2d_24 (Conv2D) (None, 10, 21, 96) 9312 multiply_4[0][0]
__________________________________________________________________________________________________
conv2d_39 (Conv2D) (None, 12, 23, 96) 9312 multiply_7[0][0]
__________________________________________________________________________________________________
add_17 (Add) (None, 10, 21, 96) 0 conv2d_6[0][0]
conv2d_9[0][0]
__________________________________________________________________________________________________
add_20 (Add) (None, 10, 21, 96) 0 conv2d_21[0][0]
conv2d_24[0][0]
__________________________________________________________________________________________________
add_23 (Add) (None, 12, 23, 96) 0 conv2d_36[0][0]
conv2d_39[0][0]
__________________________________________________________________________________________________
conv2d_10 (Conv2D) (None, 10, 21, 48) 115248 add_17[0][0]
__________________________________________________________________________________________________
conv2d_25 (Conv2D) (None, 10, 21, 48) 115248 add_20[0][0]
__________________________________________________________________________________________________
conv2d_40 (Conv2D) (None, 12, 23, 48) 115248 add_23[0][0]
__________________________________________________________________________________________________
batch_normalization_4 (BatchNor (None, 10, 21, 48) 192 conv2d_10[0][0]
__________________________________________________________________________________________________
batch_normalization_10 (BatchNo (None, 10, 21, 48) 192 conv2d_25[0][0]
__________________________________________________________________________________________________
batch_normalization_16 (BatchNo (None, 12, 23, 48) 192 conv2d_40[0][0]
__________________________________________________________________________________________________
activation_2 (Activation) (None, 10, 21, 48) 0 batch_normalization_4[0][0]
__________________________________________________________________________________________________
activation_5 (Activation) (None, 10, 21, 48) 0 batch_normalization_10[0][0]
__________________________________________________________________________________________________
activation_8 (Activation) (None, 12, 23, 48) 0 batch_normalization_16[0][0]
__________________________________________________________________________________________________
batch_normalization_5 (BatchNor (None, 10, 21, 48) 192 activation_2[0][0]
__________________________________________________________________________________________________
batch_normalization_11 (BatchNo (None, 10, 21, 48) 192 activation_5[0][0]
__________________________________________________________________________________________________
batch_normalization_17 (BatchNo (None, 12, 23, 48) 192 activation_8[0][0]
__________________________________________________________________________________________________
conv2d_11 (Conv2D) (None, 10, 21, 48) 2352 batch_normalization_5[0][0]
__________________________________________________________________________________________________
conv2d_26 (Conv2D) (None, 10, 21, 48) 2352 batch_normalization_11[0][0]
__________________________________________________________________________________________________
conv2d_41 (Conv2D) (None, 12, 23, 48) 2352 batch_normalization_17[0][0]
__________________________________________________________________________________________________
conv2d_12 (Conv2D) (None, 10, 21, 48) 20784 conv2d_11[0][0]
__________________________________________________________________________________________________
conv2d_13 (Conv2D) (None, 10, 21, 48) 20784 conv2d_11[0][0]
__________________________________________________________________________________________________
conv2d_27 (Conv2D) (None, 10, 21, 48) 20784 conv2d_26[0][0]
__________________________________________________________________________________________________
conv2d_28 (Conv2D) (None, 10, 21, 48) 20784 conv2d_26[0][0]
__________________________________________________________________________________________________
conv2d_42 (Conv2D) (None, 12, 23, 48) 20784 conv2d_41[0][0]
__________________________________________________________________________________________________
conv2d_43 (Conv2D) (None, 12, 23, 48) 20784 conv2d_41[0][0]
__________________________________________________________________________________________________
multiply_2 (Multiply) (None, 10, 21, 48) 0 conv2d_12[0][0]
conv2d_13[0][0]
__________________________________________________________________________________________________
multiply_5 (Multiply) (None, 10, 21, 48) 0 conv2d_27[0][0]
conv2d_28[0][0]
__________________________________________________________________________________________________
multiply_8 (Multiply) (None, 12, 23, 48) 0 conv2d_42[0][0]
conv2d_43[0][0]
__________________________________________________________________________________________________
conv2d_14 (Conv2D) (None, 10, 21, 48) 2352 multiply_2[0][0]
__________________________________________________________________________________________________
conv2d_29 (Conv2D) (None, 10, 21, 48) 2352 multiply_5[0][0]
__________________________________________________________________________________________________
conv2d_44 (Conv2D) (None, 12, 23, 48) 2352 multiply_8[0][0]
__________________________________________________________________________________________________
add_18 (Add) (None, 10, 21, 48) 0 conv2d_11[0][0]
conv2d_14[0][0]
__________________________________________________________________________________________________
add_21 (Add) (None, 10, 21, 48) 0 conv2d_26[0][0]
conv2d_29[0][0]
__________________________________________________________________________________________________
add_24 (Add) (None, 12, 23, 48) 0 conv2d_41[0][0]
conv2d_44[0][0]
__________________________________________________________________________________________________
global_average_pooling2d_3 (Glo (None, 48) 0 add_18[0][0]
__________________________________________________________________________________________________
global_average_pooling2d_4 (Glo (None, 48) 0 add_21[0][0]
__________________________________________________________________________________________________
global_average_pooling2d_5 (Glo (None, 48) 0 add_24[0][0]
__________________________________________________________________________________________________
tf.concat (TFOpLambda) (None, 144) 0 global_average_pooling2d_3[0][0]
global_average_pooling2d_4[0][0]
global_average_pooling2d_5[0][0]
__________________________________________________________________________________________________
dropout (Dropout) (None, 144) 0 tf.concat[0][0]
__________________________________________________________________________________________________
dense (Dense) (None, 192) 27840 dropout[0][0]
__________________________________________________________________________________________________
dropout_1 (Dropout) (None, 192) 0 dense[0][0]
__________________________________________________________________________________________________
dense_1 (Dense) (None, 24) 4632 dropout_1[0][0]
==================================================================================================
Total params: 41,756,881
Trainable params: 41,737,483
Non-trainable params: 19,398
__________________________________________________________________________________________________
@tf.function
def _mixup(inp, targ):
indice = tf.range(len(inp))
indice = tf.random.shuffle(indice)
sinp = tf.gather(inp, indice, axis=0)
starg = tf.gather(targ, indice, axis=0)
alpha = 0.2
t = tf.compat.v1.distributions.Beta(alpha, alpha).sample([len(inp)])
tx = tf.reshape(t, [-1, 1, 1, 1])
ty = tf.reshape(t, [-1, 1])
x = inp * tx + sinp * (1-tx)
y = targ * ty + starg * (1-ty)
# y = tf.minimum(targ + starg, 1.0) # for multi-label???
return x, y
tfrecs = sorted(tf.io.gfile.glob(TRAIN_TFREC + '/*.tfrec'))
parsed_trainval = (tf.data.TFRecordDataset(tfrecs, num_parallel_reads=AUTOTUNE)
.map(_parse_function, num_parallel_calls=AUTOTUNE).unbatch()
.filter(_filtTP).enumerate())
indices = []
spid = []
recid = []
for i, sample in tqdm(parsed_trainval.prefetch(AUTOTUNE)):
indices.append(i.numpy())
spid.append(sample['species_id'].numpy())
recid.append(sample['recording_id'].numpy().decode())
1216it [00:36, 33.18it/s]
table = pd.DataFrame({'indices': indices, 'species_id': spid, 'recording_id': recid})
table
| indices | species_id | recording_id | |
|---|---|---|---|
| 0 | 0 | 14 | 003bec244 |
| 1 | 1 | 12 | 2026bced7 |
| 2 | 2 | 21 | 422de4e4d |
| 3 | 3 | 6 | 60a493ad4 |
| 4 | 4 | 13 | 8080b2283 |
| ... | ... | ... | ... |
| 1211 | 1211 | 3 | 807efd6bb |
| 1212 | 1212 | 20 | a6610076b |
| 1213 | 1213 | 23 | c91cae4aa |
| 1214 | 1214 | 3 | c91cae4aa |
| 1215 | 1215 | 1 | e755e15ec |
1216 rows × 3 columns
skf = StratifiedKFold(n_splits=5, random_state=SEED, shuffle=True)
splits = list(skf.split(table.index, table.species_id))
plt.hist([table.loc[splits[0][0], 'species_id'], table.loc[splits[0][1], 'species_id']], bins=CLASS_N,stacked=True)
plt.show()
def create_idx_filter(indice):
@tf.function
def _filt(i, x):
return tf.reduce_any(indice == i)
return _filt
@tf.function
def _remove_idx(i, x):
return x
def create_train_dataset(batchsize, train_idx):
global parsed_trainval
parsed_train = (parsed_trainval
.filter(create_idx_filter(train_idx))
.map(_remove_idx))
dataset = (parsed_train.cache()
.shuffle(len(train_idx))
.repeat()
.map(_cut_wav, num_parallel_calls=AUTOTUNE)
.map(_wav_to_spec, num_parallel_calls=AUTOTUNE)
.map(_create_annot, num_parallel_calls=AUTOTUNE)
.map(_preprocess, num_parallel_calls=AUTOTUNE)
.batch(batchsize))
if cfg['model_params']['mixup']:
dataset = (dataset.map(_mixup, num_parallel_calls=AUTOTUNE)
.prefetch(AUTOTUNE))
else:
dataset = dataset.prefetch(AUTOTUNE)
return dataset
def create_val_dataset(batchsize, val_idx):
global parsed_trainval
parsed_val = (parsed_trainval
.filter(create_idx_filter(val_idx))
.map(_remove_idx))
vdataset = (parsed_val
.map(_cut_wav_val, num_parallel_calls=AUTOTUNE)
.map(_wav_to_spec, num_parallel_calls=AUTOTUNE)
.map(_create_annot, num_parallel_calls=AUTOTUNE)
.map(_preprocess_val, num_parallel_calls=AUTOTUNE)
.batch(8*tpu_strategy.num_replicas_in_sync)
.cache())
return vdataset
# from https://www.kaggle.com/carlthome/l-lrap-metric-for-tf-keras
@tf.function
def _one_sample_positive_class_precisions(example):
y_true, y_pred = example
retrieved_classes = tf.argsort(y_pred, direction='DESCENDING')
class_rankings = tf.argsort(retrieved_classes)
retrieved_class_true = tf.gather(y_true, retrieved_classes)
retrieved_cumulative_hits = tf.math.cumsum(tf.cast(retrieved_class_true, tf.float32))
idx = tf.where(y_true)[:, 0]
i = tf.boolean_mask(class_rankings, y_true)
r = tf.gather(retrieved_cumulative_hits, i)
c = 1 + tf.cast(i, tf.float32)
precisions = r / c
dense = tf.scatter_nd(idx[:, None], precisions, [y_pred.shape[0]])
return dense
class LWLRAP(tf.keras.metrics.Metric):
def __init__(self, num_classes, name='lwlrap'):
super().__init__(name=name)
self._precisions = self.add_weight(
name='per_class_cumulative_precision',
shape=[num_classes],
initializer='zeros',
)
self._counts = self.add_weight(
name='per_class_cumulative_count',
shape=[num_classes],
initializer='zeros',
)
def update_state(self, y_true, y_pred, sample_weight=None):
precisions = tf.map_fn(
fn=_one_sample_positive_class_precisions,
elems=(y_true, y_pred),
dtype=(tf.float32),
)
increments = tf.cast(precisions > 0, tf.float32)
total_increments = tf.reduce_sum(increments, axis=0)
total_precisions = tf.reduce_sum(precisions, axis=0)
self._precisions.assign_add(total_precisions)
self._counts.assign_add(total_increments)
def result(self):
per_class_lwlrap = self._precisions / tf.maximum(self._counts, 1.0)
per_class_weight = self._counts / tf.reduce_sum(self._counts)
overall_lwlrap = tf.reduce_sum(per_class_lwlrap * per_class_weight)
return overall_lwlrap
def reset_states(self):
self._precisions.assign(self._precisions * 0)
self._counts.assign(self._counts * 0)
def _parse_function_test(example_proto):
sample = tf.io.parse_single_example(example_proto, feature_description)
wav, _ = tf.audio.decode_wav(sample['audio_wav'], desired_channels=1) # mono
@tf.function
def _cut_audio(i):
_sample = {
'audio_wav': tf.reshape(wav[i*SR*TIME:(i+1)*SR*TIME], [SR*TIME]),
'recording_id': sample['recording_id']
}
return _sample
return tf.map_fn(_cut_audio, tf.range(60//TIME), dtype={
'audio_wav': tf.float32,
'recording_id': tf.string
})
def inference(model):
tdataset = (tf.data.TFRecordDataset(tf.io.gfile.glob(TEST_TFREC + '/*.tfrec'), num_parallel_reads=AUTOTUNE)
.map(_parse_function_test, num_parallel_calls=AUTOTUNE).unbatch()
.map(_wav_to_spec, num_parallel_calls=AUTOTUNE)
.map(_preprocess_test, num_parallel_calls=AUTOTUNE)
.batch(128*(60//TIME)).prefetch(AUTOTUNE))
rec_ids = []
probs = []
for inp, rec_id in tqdm(tdataset):
with tpu_strategy.scope():
pred = model.predict_on_batch(tf.reshape(inp, [-1, HEIGHT, WIDTH, 3]))
prob = tf.sigmoid(pred)
prob = tf.reduce_max(tf.reshape(prob, [-1, 60//TIME, CLASS_N]), axis=1)
rec_id_stack = tf.reshape(rec_id, [-1, 60//TIME])
for rec in rec_id.numpy():
assert len(np.unique(rec)) == 1
rec_ids.append(rec_id_stack.numpy()[:,0])
probs.append(prob.numpy())
crec_ids = np.concatenate(rec_ids)
cprobs = np.concatenate(probs)
sub = pd.DataFrame({
'recording_id': list(map(lambda x: x.decode(), crec_ids.tolist())),
**{f's{i}': cprobs[:,i] for i in range(CLASS_N)}
})
sub = sub.sort_values('recording_id')
return sub
def plot_history(history, name):
plt.figure(figsize=(8,3))
plt.subplot(1,2,1)
plt.plot(history.history["loss"])
plt.plot(history.history["val_loss"])
plt.legend(['Train', 'Test'], loc='upper left')
plt.title("loss")
# plt.yscale('log')
plt.subplot(1,2,2)
plt.plot(history.history["lwlrap"])
plt.plot(history.history["val_lwlrap"])
plt.legend(['Train', 'Test'], loc='upper left')
plt.title("metric")
plt.savefig(name)
def train_and_inference(splits, split_id):
print(split_id)
batchsize = cfg['model_params']['batchsize_per_tpu'] * tpu_strategy.num_replicas_in_sync
print("batchsize", batchsize)
loss_fn = cfg['model_params']['loss']['fn'](from_logits=True, **cfg['model_params']['loss']['params'])
idx_train_tf = tf.constant(splits[split_id][0])
idx_val_tf = tf.constant(splits[split_id][1])
dataset = create_train_dataset(batchsize, idx_train_tf)
vdataset = create_val_dataset(batchsize, idx_val_tf)
optimizer = cfg['model_params']['optim']['fn'](**cfg['model_params']['optim']['params'])
with tpu_strategy.scope():
model = create_model()
model.compile(optimizer=optimizer, loss=loss_fn, metrics=[LWLRAP(CLASS_N)])
if split_id not in (10,10):#!!!!!!!!!!!!! For convenience: If your Colab shuts down for some reason, you can always download your already trained models from Google Drive
history = model.fit(dataset,
steps_per_epoch=cfg['model_params']['iteration_per_epoch'],
epochs=cfg['model_params']['epoch'],
validation_data=vdataset,
callbacks=[
tf.keras.callbacks.ReduceLROnPlateau(
'val_lwlrap', patience=10
),
tf.keras.callbacks.ModelCheckpoint(
filepath=models_path+'model_best_%d.h5' % split_id,
save_weights_only=True,
monitor='val_lwlrap',
mode='max',
save_best_only=True),
]
)
plot_history(history, 'history_%d.png' % split_id)
best_score = max(history.history['val_lwlrap'])
print (best_score)
### inference ###
model.load_weights(models_path+'model_best_%d.h5' % split_id)
sub=inference(model)
del model
gc.collect()
return sub,best_score
# train and inference
# sub, _ = train_and_inference(splits, 0)
# N-fold ensemble
""" Delete this line to start training the model
print(SEED)
train_n=0
df = pd.DataFrame(columns=["train_n",'split_id','best_score','CSV','SEED'])
for split_id in range(len(splits)):
sub, best_score=train_and_inference(splits, split_id)
sub.set_index('recording_id').to_csv(models_path+f"submission_train_n_{train_n}_split_id_{split_id}.csv", index=False)
df = df.append({'train_n': train_n,'split_id': split_id,'best_score': best_score,'CSV': f"submission_train_n_{train_n}_split_id_{split_id}.csv",'SEED':SEED}, ignore_index=True)
df.to_csv(models_path+f"train_n_{train_n}.csv", index=False)
#"""
' Delete this line to start training the model\nprint(SEED)\ntrain_n=0\ndf = pd.DataFrame(columns=["train_n",\'split_id\',\'best_score\',\'CSV\',\'SEED\'])\nfor split_id in range(len(splits)):\n sub, best_score=train_and_inference(splits, split_id)\n sub.set_index(\'recording_id\').to_csv(models_path+f"submission_train_n_{train_n}_split_id_{split_id}.csv", index=False)\n df = df.append({\'train_n\': train_n,\'split_id\': split_id,\'best_score\': best_score,\'CSV\': f"submission_train_n_{train_n}_split_id_{split_id}.csv",\'SEED\':SEED}, ignore_index=True)\ndf.to_csv(models_path+f"train_n_{train_n}.csv", index=False)\n#'
#sub.describe()
If you have questions then ask, I will help as I can